Establishment of an algorithm for automated detection of iPS/non-iPS cells
under a culture condition by noninvasive image analysis
Hirotada Watanabe1, Koji Tanabe2, Hiroaki Kii1, Momotaro Ishikawa1, Chieko Nakada1, Takayuki Uozumi1, Yasujiro Kiyota1, Youichi Wada1, Ryoji Tsuchiya1
1
2
Instruments company, NIKON CORPORATION, Yokohama, Japan, Center for iPS Cell Research and Application (CiRA), Kyoto University, Kyoto, Japan
Introduction
Algorithm development
Algorithm validation
There are great expectations for iPS cells as tools for drug development and resources for
regenerative medicine. However, such medical applications demand a stable supply of iPS cells
in large quantities and many issues for optimization of iPS cell generation remain to be resolved.
In addition, evaluation of iPS cells is time-consuming and dependent on the skills of individual
technicians.
We have developed a system that can observe a large quantity of iPS cells automatically and
classify iPS colonies with a standardized algorithm.
In this experiment, a cell culture observation system (BioStation CT) was used to observe a
culture, and image analysis software (CL-Quant) was used to scan phase-contrast images and
automatically detect iPS/non-iPS colonies generated by reprogramming of normal somatic
cells. In implementing the automatic detection an algorithm was designed to categorize cells of
different origin based on parameters pertaining to morphological characteristics of the cells. For
20 samples consisting of various numbers of colonies the coefficient of correlation was extremely
high between the number of iPS colonies counted by inspection and the number of iPS colonies
counted by CL-Quant. Accuracy in iPS/non-iPS identification was next compared. For the cell type
used for algorithm design, accuracy was high at an average of 79.9%. For three types of cells
of different origin the average accuracy was 80.3%. This indicates the strength of the method
in providing accurate iPS/non-iPS identification results. To confirm the consistency of analysis
results, accuracy was compared with the results of analyzing cells stained with the TRA-1-60
antibody (marker specific to iPS cells). This experiment shows that iPS/non-iPS identification
during the establishment stage can be automated eliminating the need for human intervention
and allowing for non-invasive analysis without the use of fluorescent dyes. This provides several
advantages. 1) Reduction of required time and effort. 2) Non-invasive analysis.
3) Homogenization. The method allows for consistent selection of good iPS cells without
requiring training. 4) Historical management. By maintaining a record of the identification results,
quality evaluations can be reanalyzed with ease.
Step 1) Total colony detection
1) iPS colony count evaluation
iPS cell reprogramming
Image acquisition
Image analysis
Original image
Detection of masks for colony by machine
learning technology (Soft MatchingTM)
Flatten Background
Fig. 8 shows plotted results of manual and CL-Quant-made iPS counts of iPS colonies (in 20 samples)
used for creating the algorithm. The linear approximation of the graph shows a high correlation function
of R=0.95. This indicates that CL-Quant-made iPS counts have similar results to manual counts of any
number of colonies. The difference between target areas of manual and CL-Quant-made counts may be
less than 1 linear approximation slope.
400
iPS count by CL-Quant
350
colony
background
Partition (large colony)
Removing outside region
y = 0.6756x + 20.772
R = 0.9099
200
150
100
50
0
iPS
non-iPS
Fig. 7 Analysis result image
0
100
200
300
400
500
600
iPS count by human
Fig. 8 Comparison of manual and CL-Quant-made iPS colony counts
2) Accuracy evaluation for different cell lines
Compare accuracies of manual and CL-Quant-made iPS/non-iPS identifications using iPS cell (Cell
A), used for creating an algorithm, and other cell lines (Cell B, C, D). CL-Quant-made iPS/non-iPS
identification using Cell A indicates an accuracy of 79.9%, and other cell lines 80.3%. This shows the
strength of this analysis algorithm.
Fig. 3 Total colony detection step flow
Step 2) iPS colony identification
iPS colony identification steps are shown in Fig. 4. Colonies detected in Step 1) are separated by size.
Additional colony boundary masks are given to each colony (Fig. 5). Each colony is then classified as iPS
or non-iPS based on the morphological rules depicted by the decision tree. The decision tree was created
by teaching iPS/non-iPS colonies to CL-Quant software (Fig.6). Classified colonies are referred to the
total iPS/non-iPS colony number in the vessel.
Create colony boundary
mask (erode and
subtract)
Separate colonies by
size (150μm)
iPS cell reprogramming
Type
A
B
C
D
Large
Original
Human
iPS
non-iPS
iPS
non-iPS
80
CL-Quant
iPS
non-iPS
non-iPS
iPS
70
60
50
40
30
20
10
0
Cell A
Cell B
Cell C
Cell D
Cell line
Fig. 9 Accuracy evaluation of iPS/non-iPS identification
Left: Definition of Accuracy.
Right: Accuracy comparison of iPS/non-iPS identification of several cell lines (use dozens of colonies for 1 data, N=3).
iPS/non-iPS
classification by large
colony decision tree
Colony mask
90
A+B
Accuracy (%)=
X100
A+B+C+D
3) Comparison with iPS cell specific marker
Comparison results of TRA-1-60 positive/negative identification and CL-Quant-made iPS/non-iPS
identification using TRA-1-60 immunofluorescent stain (an iPS cell specific marker).
Comparing results of TRA-1-60 stain and CL-Quant-made identifications of 237 colonies, the results
coincide for 70.4% of colonies (167 colonies), indicating the high consistency of both identifications.
(Fig. 10, Table 1)
Fig. 4 iPS colony identification step flow
iPS cells were generated from adult human dermal fibroblast(HDF) by retroviral transduction of four factors: Oct3/4, Sox2, Klf4, and c-Myc.[1]
Six days after transduction, the cells were harvested by trypsinization and plated onto mitomycin C-treated SNL feeder cells using 100 mm dish.
100
Accuracy
iPS/non-iPS
classification by small
colony decision tree
Total count of iPS/noniPS colonies
Create colony boundary
mask (erode and
subtract)
Fig. 1 Experimental step flow of iPS cell reprogramming, Image acquisition, and Image analysis
Colony boundary mask
Type A
Type B
Type C
Type D
iPS
Image acquisition
Phase contrast images and fluorescence images of 100 mm dish were acquired using BioStation CT (NIKON CORPORATION) at 37°C, 5% CO2.
Full Scan mode (2x) was set for acquiring full well tiling images.
Tiling mode (10x) was set for acquiring partial well tiling images.
250
Total colony
Small
Oct3/4, Sox2, Klf4, c-myc
300
Accuracy (%)
Methods
To detect colonies accurately and quickly from stitched images, original stitched images are preprocessed and colony regions are recognized as colony masks using machine leaning technology
(Soft MatchingTM). Created colony masks undergo post-processing as follows (Fig. 3).
Fig. 10 TRA-1-60 immunofluorescent stained colony images (10x tiling images)
Non-iPS
Table 1 Comparison of TRA-1-60 immunofluorescent stain result with CL-Quant iPS/non-iPS identification
Fig. 5 Typical iPS/non-iPS colony images and created colony masks and colony boundary masks
Fig. 2 BioStation CT
BioStation CT is a cell culture observation system that
allows microscope observation of multiple vessels (max.
30) easily and automatically under culture conditions.
It is possible to acquire full-well tiling 2x images
by selecting Full Scan mode. Acquired images are
automatically downloaded onto a PC for analysis.
Compactness
Type
TRA-1-60
CL-Quant
A
+
iPS
B
-
non-iPS
C
+
non-iPS
D
Compactness
= (perimeter)2/ size
-
iPS
Percentage
65.8%
(156 colonies)
4.6%
(11 colonies)
28.7%
(68 colonies)
0.8%
(2 colonies)
Accuracy
○ 70.4%
× 29.5%
Reference
Round
Indent
1. Takahashi, K., Tanabe, K., Ohnuki, M., Narita, M., Ichisaka, T., Tomoda, K. & Yamanaka, S. 2007
Induction of Pluripotent Stem Cells from Adult Human Fibroblasts by Defined Factors. Cell 131, 861-872.
2. Alworth, SV., Watanabe, H. & Lee, JSJ. 2010 Teachable, High-Content Analytics for Live-Cell, Phase
Contrast Movies. J Biomol Screening 8, 968-977.
Conclusion
Image analysis
Full well tiling images were loaded into CL-Quant software (NIKON CORPORATION) and stitched to create 18,000-by-18,000 pixel composite images for analysis.
All image analyses (colony segmentation, measurement, colony classification etc.) were performed using functions built into CL-Quant software [2].
Fig. 6 Classification rule.
Left: A portion of a decision tree. 2695 colonies were taught (652 iPS, 2043 non-iPS) to create the small
colony decision tree. 2013 colonies were taught (899 iPS, 1114 non-iPS) to create the large colony
decision tree. Decision trees contain a variety of colony morphological and intensity information (size,
intensity of boundary region, compactness, etc.) In both trees, the compactness of the colony mask is
the primary rule (first decision node). This shows that iPS colonies have more rounded colony boundaries
than non-iPS colonies.
Right: The compactness of the object.
• Nikon has developed an image analysis algorithm that detects colonies in the entire area of a dish and
identifies iPS/non-iPS colonies using certain standards, such as compactness.
• iPS/non-iPS identification using the analysis algorithm shows similar results to manual identification
(R=0.95) and also shows similar results (70.4% of agreement rate) to identification using iPS cell specific
markers. In addition, the algorithm proves it is highly accurate (80.3%) for different cell lines.
• These noninvasive analysis systems can automatically acquire a number of images of samples and
analyze images based on certain standards. These functions reduce variations in evaluation during iPS
cell generation and save time and work.